SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC On-Demand

Presentation
Media
Abstract:
We will discuss a deep learning-based method for improving the quality of 3D reconstruction performed by time-of-flight cameras. Scene motion, multiple reflections, and sensor noise introduce artifacts in the depth reconstruction performed by these sensors. We'll explain our proposed two-stage, deep-learning approach to address all of these sources of artifacts simultaneously. We'll also introduce FLAT, a synthetic dataset of 2000 ToF measurements that capture all of these nonidealities and can be used to simulate different hardware. Using the Kinect camera as a baseline, we show improved reconstruction errors on simulated and real data, as compared with state-of-the-art methods.
We will discuss a deep learning-based method for improving the quality of 3D reconstruction performed by time-of-flight cameras. Scene motion, multiple reflections, and sensor noise introduce artifacts in the depth reconstruction performed by these sensors. We'll explain our proposed two-stage, deep-learning approach to address all of these sources of artifacts simultaneously. We'll also introduce FLAT, a synthetic dataset of 2000 ToF measurements that capture all of these nonidealities and can be used to simulate different hardware. Using the Kinect camera as a baseline, we show improved reconstruction errors on simulated and real data, as compared with state-of-the-art methods.  Back
 
Topics:
AI and DL Research, Computer Vision
Type:
Talk
Event:
GTC Silicon Valley
Year:
2019
Session ID:
S9318
Streaming:
Download:
Share:
 
Abstract:
Training intelligent agents with reinforcement learning is a notoriously unstable process. Although massive parallelization on GPUs and distributed systems can reduce instabilities, the success of training remains strongly influenced by the choice of hyperparameters. We'll describe a novel meta-optimization algorithm for distributed systems that solves a set of optimization problems in parallel while looking for the optimal hyperparameters. We'll also show how it applies to deep reinforcement learning. We'll demonstrate how the algorithm can fine-tune hyperparameters while learning to play different Atari games. Compared with existing approaches, our algorithm releases more computational resources during training by means of a stochastic scheduling procedure. Our algorithm has been implemented on top of MagLev, the NVIDIA AI training and inference infrastructure.
Training intelligent agents with reinforcement learning is a notoriously unstable process. Although massive parallelization on GPUs and distributed systems can reduce instabilities, the success of training remains strongly influenced by the choice of hyperparameters. We'll describe a novel meta-optimization algorithm for distributed systems that solves a set of optimization problems in parallel while looking for the optimal hyperparameters. We'll also show how it applies to deep reinforcement learning. We'll demonstrate how the algorithm can fine-tune hyperparameters while learning to play different Atari games. Compared with existing approaches, our algorithm releases more computational resources during training by means of a stochastic scheduling procedure. Our algorithm has been implemented on top of MagLev, the NVIDIA AI training and inference infrastructure.  Back
 
Topics:
Deep Learning and AI Frameworks, HPC and AI
Type:
Talk
Event:
GTC Silicon Valley
Year:
2019
Session ID:
S9414
Streaming:
Download:
Share:
 
Abstract:
The paradigm for robot programming is changing with the adoption of the deep learning approach in the field of robotics. Instead of hard coding a complex sequence of actions, tasks are acquired by the robot through an active learning procedure. This introduces new challenges that have to be solved to achieve effective training. We'll show several issues that can be encountered while learning a close-loop DNN controller aimed at a fundamental task like grasping, and their practical solutions. First, we'll illustrate the advantages of training using a simulator, as well as the effects of choosing different learning algorithms in the reinforcement learning and imitation learning domains. We'll then show how separating the control and vision modules in the DNN can simplify and speed up the learning procedure in the simulator, although the learned controller hardly generalizes to the real world environment. Finally, we'll demonstrate how to use domain transfer to train a DNN controller in a simulator that can be effectively employed to control a robot in the real world.
The paradigm for robot programming is changing with the adoption of the deep learning approach in the field of robotics. Instead of hard coding a complex sequence of actions, tasks are acquired by the robot through an active learning procedure. This introduces new challenges that have to be solved to achieve effective training. We'll show several issues that can be encountered while learning a close-loop DNN controller aimed at a fundamental task like grasping, and their practical solutions. First, we'll illustrate the advantages of training using a simulator, as well as the effects of choosing different learning algorithms in the reinforcement learning and imitation learning domains. We'll then show how separating the control and vision modules in the DNN can simplify and speed up the learning procedure in the simulator, although the learned controller hardly generalizes to the real world environment. Finally, we'll demonstrate how to use domain transfer to train a DNN controller in a simulator that can be effectively employed to control a robot in the real world.  Back
 
Topics:
AI and DL Research, IoT, Robotics & Drones, Computer Vision, Robotics & Autonomous Machines
Type:
Talk
Event:
GTC Silicon Valley
Year:
2018
Session ID:
S8132
Streaming:
Download:
Share:
 
Abstract:
Traditional RL training is dominated by experience collection processes executing on the CPU. However, this CPU oriented design pattern limits the utility of DL accelerators, such as GPUs. In this talk we present CuLE (cuda learning environment), an experimental deep RL companion library, to facilitate the generation of RL updates directly on the GPU. CuLE provides an implementation of ALE (atari learning environment), a challenging RL benchmark for discrete episodic tasks, executing directly on the GPU with the number of environments ranging from a few hundred to several thousand. Although traditional deep RL implementations use 12-16 agents coupled with replay memory to achieve training efficiency CuLE can generate a massive number of samples per step and supports new training scenarios that minimize expensive data movement operations. With 1024 agents CuLE achieves an 8-10x performance improvement by executing directly on the GPU compared to 1024 agents running in parallel on a 12-core CPU. We plan to extend CuLE to support a new set GPU-centric deep RL training schemes and new challenging training environments through integration with GFN.?
Traditional RL training is dominated by experience collection processes executing on the CPU. However, this CPU oriented design pattern limits the utility of DL accelerators, such as GPUs. In this talk we present CuLE (cuda learning environment), an experimental deep RL companion library, to facilitate the generation of RL updates directly on the GPU. CuLE provides an implementation of ALE (atari learning environment), a challenging RL benchmark for discrete episodic tasks, executing directly on the GPU with the number of environments ranging from a few hundred to several thousand. Although traditional deep RL implementations use 12-16 agents coupled with replay memory to achieve training efficiency CuLE can generate a massive number of samples per step and supports new training scenarios that minimize expensive data movement operations. With 1024 agents CuLE achieves an 8-10x performance improvement by executing directly on the GPU compared to 1024 agents running in parallel on a 12-core CPU. We plan to extend CuLE to support a new set GPU-centric deep RL training schemes and new challenging training environments through integration with GFN.?  Back
 
Topics:
Deep Learning and AI Frameworks, Tools and Libraries
Type:
Talk
Event:
GTC Silicon Valley
Year:
2018
Session ID:
S8440
Streaming:
Download:
Share:
 
Abstract:
We'll introduce a hybrid CPU/GPU version of the Asynchronous Advantage Actor-Critic (A3C) algorithm, currently the state-of-the-art method in reinforcement learning for various gaming tasks. We'll analyze its computational traits and concentrate on the critical aspects to leverage the GPU's computational power. We'll introduce a system of queues and a dynamic scheduling strategy, potentially helpful for other asynchronous algorithms as well. Our hybrid CPU/GPU version of A3C, based on TensorFlow, achieves a significant speed-up compared to a CPU implementation and is publicly available to other researchers.
We'll introduce a hybrid CPU/GPU version of the Asynchronous Advantage Actor-Critic (A3C) algorithm, currently the state-of-the-art method in reinforcement learning for various gaming tasks. We'll analyze its computational traits and concentrate on the critical aspects to leverage the GPU's computational power. We'll introduce a system of queues and a dynamic scheduling strategy, potentially helpful for other asynchronous algorithms as well. Our hybrid CPU/GPU version of A3C, based on TensorFlow, achieves a significant speed-up compared to a CPU implementation and is publicly available to other researchers.  Back
 
Topics:
Deep Learning and AI, Performance Optimization, Game Development
Type:
Talk
Event:
GTC Silicon Valley
Year:
2017
Session ID:
S7169
Download:
Share: